What are the differences between feather and parquet?

Question

Both are columnar (disk-)storage formats for use in data analysis systems. Both are integrated within Apache Arrow (pyarrow package for python) and are designed to correspond with Arrow as a columnar in-memory analytics layer.

How do both formats differ?

Should you always prefer feather when working with pandas when possible?

What are the use cases where feather is more suitable than parquet and the other way round?

---

Appendix

I found some hints here https://github.com/wesm/feather/issues/188, but given the young age of this project, it's possibly a bit out of date.

Not a serious speed test because I'm just dumping and loading a whole Dataframe but to give you some impression if you never heard of the formats before:

 # IPython    
import numpy as np
import pandas as pd
import pyarrow as pa
import pyarrow.feather as feather
import pyarrow.parquet as pq
import fastparquet as fp


df = pd.DataFrame({'one': [-1, np.nan, 2.5],
                   'two': ['foo', 'bar', 'baz'],
                   'three': [True, False, True]})

print("pandas df to disk ####################################################")
print('example_feather:')
%timeit feather.write_feather(df, 'example_feather')
# 2.62 ms ± 35.8 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
print('example_parquet:')
%timeit pq.write_table(pa.Table.from_pandas(df), 'example.parquet')
# 3.19 ms ± 51 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
print()

print("for comparison:")
print('example_pickle:')
%timeit df.to_pickle('example_pickle')
# 2.75 ms ± 18.8 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
print('example_fp_parquet:')
%timeit fp.write('example_fp_parquet', df)
# 7.06 ms ± 205 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)
print('example_hdf:')
%timeit df.to_hdf('example_hdf', 'key_to_store', mode='w', table=True)
# 24.6 ms ± 4.45 ms per loop (mean ± std. dev. of 7 runs, 100 loops each)
print()

print("pandas df from disk ##################################################")
print('example_feather:')
%timeit feather.read_feather('example_feather')
# 969 µs ± 1.8 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
print('example_parquet:')
%timeit pq.read_table('example.parquet').to_pandas()
# 1.9 ms ± 5.5 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

print("for comparison:")
print('example_pickle:')
%timeit pd.read_pickle('example_pickle')
# 1.07 ms ± 6.21 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
print('example_fp_parquet:')
%timeit fp.ParquetFile('example_fp_parquet').to_pandas()
# 4.53 ms ± 260 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)
print('example_hdf:')
%timeit pd.read_hdf('example_hdf')
# 10 ms ± 43.4 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

# pandas version: 0.22.0
# fastparquet version: 0.1.3
# numpy version: 1.13.3
# pandas version: 0.22.0
# pyarrow version: 0.8.0
# sys.version: 3.6.3
# example Dataframe taken from https://arrow.apache.org/docs/python/parquet.html

Answer 1

• Parquet format is designed for long-term storage, where Arrow is more intended for short term or ephemeral storage (Arrow may be more suitable for long-term storage after the 1.0.0 release happens, since the binary format will be stable then)

• Parquet is more expensive to write than Feather as it features more layers of encoding and compression. Feather is unmodified raw columnar Arrow memory. We will probably add simple compression to Feather in the future.

• Due to dictionary encoding, RLE encoding, and data page compression, Parquet files will often be much smaller than Feather files

• Parquet is a standard storage format for analytics that's supported by many different systems: Spark, Hive, Impala, various AWS services, in future by BigQuery, etc. So if you are doing analytics, Parquet is a good option as a reference storage format for query by multiple systems

The benchmarks you showed are going to be very noisy since the data you read and wrote is very small. You should try compressing at least 100MB or upwards 1GB of data to get some more informative benchmarks, see e.g. http://wesmckinney.com/blog/python-parquet-multithreading/

Hope this helps